Anthropic tool-format adapter & middleware

[dgenio]

Parent: #40 | Split: Anthropic adapter

Problem

This is part 2 of #40. Developers using Anthropic's API must manually translate between Capability objects and Anthropic's {name, description, input_schema} tool schema with cache_control support.

Proposed Change

1. Schema adapter (src/agent_kernel/adapters/anthropic.py)

def capabilities_to_tools(caps: list[Capability]) -> list[dict]:
    """Convert Capabilities to Anthropic tool definitions."""

def tool_use_to_request(tool_use_block: dict) -> CapabilityRequest:
    """Convert an Anthropic tool_use block to a CapabilityRequest."""

def format_result(result: Any, tool_use_id: str) -> dict:
    """Format kernel invoke() result as Anthropic tool_result."""

2. Middleware class

class AnthropicMiddleware:
    def __init__(self, kernel: Kernel, principal: Principal): ...
    def get_tools(self) -> list[dict]: ...
    async def handle_tool_uses(self, content_blocks: list[dict]) -> list[dict]: ...
    def intercept_tool_call(self, callback: Callable) -> None: ...
    def intercept_tool_result(self, callback: Callable) -> None: ...

3. No SDK dependency

Adapters work with plain dicts — no dependency on anthropic package.

Acceptance Criteria

• Round-trip: Capability → Anthropic schema → tool_use → CapabilityRequest preserves all fields
• AnthropicMiddleware.handle_tool_uses() processes through full kernel pipeline
• cache_control support in tool definitions
• Hooks execute in correct order
• Schema matches current Anthropic API spec
• No runtime dependency on anthropic package

Affected Files

• src/agent_kernel/adapters/anthropic.py (new)
• tests/test_adapters.py (extend)
• docs/integrations.md (usage examples)

Dependencies

• Can be developed independently or in parallel with the OpenAI adapter